Abstract: Abstract: Road water accumulation is a widespread traffic safety hazard that can easily lead to vehicle loss of control and cause traffic accidents ranging from minor scrapes to serious collisions. Accurately identifying road water accumulation is hampered by challenges such as complex road surface textures, variable lighting conditions, and the diverse morphology of the water itself. To address these issues, this study first designs the RAFE reflectivity-aware feature enhancement module, which effectively integrates shallow texture and deep semantic features. It also suppresses specular interference through an adaptive noise weighting mechanism, improving the model's robustness under complex lighting conditions. Secondly, the C3k2_Enhanced module is designed, combining grouped convolution with a lightweight channel attention mechanism. This reduces the number of model parameters while enhancing the ability to capture key features such as small water flows. Finally, a lightweight shared detail enhancement detection head (Detect_LSDECD) is introduced. This optimizes multi-scale feature fusion through shared convolution and learnable scale parameters, significantly improving the detection accuracy and boundary localization of small water accumulation areas. Experiments show that compared to the original YOLOv11n model, REL-YOLO significantly improves recall (R) and mean average precision (mAP@0.5) by 3.2% and 2.0%, respectively, at an IoU threshold of 0.6. It also reduces the number of model parameters by 15% and significantly reduces computational complexity (GFLOPs). While maintaining high real-time performance (FPS), REL-YOLO strikes a balance between accuracy and efficiency, providing a highly effective solution for real-time road flooding detection in complex environments.

Key words: Keywords: road ponding, object detection, reflection perception, feature enhancement, lightweight model, REL-YOLO

摘要： 摘 要: 道路积水作为一种广泛存在的交通安全隐患，易导致车辆失控，引发从轻微刮蹭到严重碰撞等不同程度的交通事故。准确识别道路积水受限于复杂路面纹理、多变光照条件及积水本身形态多样等挑战。针对上述问题，本研究首先设计RAFE反射感知特征增强模块，有效融合浅层纹理与深层语义特征，并通过自适应噪声加权机制抑制镜面反射干扰，提升模型在复杂光照下的鲁棒性。其次，设计C3k2_Enhanced模块，结合分组卷积与轻量化通道注意力机制，在降低模型参数量的同时增强了对细小水流等关键特征的捕获能力。最后，引入轻量级共享细节增强检测头(Detect_LSDECD)，通过共享卷积与可学习尺度参数优化多尺度特征融合，显著提升小目标积水区域的检测精度与边界定位能力。实验表明，相较于原始YOLOv11n模型，REL-YOLO在IoU阈值为0.6时，召回率(R)和平均精度(mAP@0.5)分别显著提升了3.2%和2.0%，同时模型参数量减少了15%，计算复杂度(GFLOPs)显著降低。REL-YOLO在保持高实时性(FPS)的同时，实现了精度与效率的平衡，为复杂环境下的道路积水实时检测提供了高效解决方案。

关键词: 关键词: 道路积水, 目标检测, 反射感知, 特征增强, 轻量化模型, REL-YOLO